Merge pull request #37 from eth-cscs/release/0.9.13

Release/0.9.13
eth-cscs · Aug 18, 2020 · b00fbbc · b00fbbc
2 parents c35e0a5 + 70c491a
commit b00fbbc
Show file tree

Hide file tree

Showing 5 changed files with 49 additions and 2 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -1,5 +1,5 @@
 cmake_minimum_required(VERSION 3.11 FATAL_ERROR) # 3.11 to avoid issues with OpenMP + CUDA
-project(SpFFT LANGUAGES CXX VERSION 0.9.12)
+project(SpFFT LANGUAGES CXX VERSION 0.9.13)
 set(SPFFT_SO_VERSION 0)
 set(SPFFT_VERSION ${PROJECT_VERSION})
 

diff --git a/cmake/modules/FindROCFFT.cmake b/cmake/modules/FindROCFFT.cmake
@@ -56,7 +56,7 @@ endif()
 
 find_library(
     ROCFFT_LIBRARIES
-    NAMES "rocfft"
+    NAMES "rocfft" "rocfft-d"
     HINTS ${_ROCFFT_PATHS}
     PATH_SUFFIXES "rocfft/lib" "rocfft" 
 )

diff --git a/src/fft/transform_real_2d_gpu.hpp b/src/fft/transform_real_2d_gpu.hpp
@@ -198,6 +198,23 @@ class TransformReal2DGPU : public TransformGPU {
   inline auto device_id() const noexcept -> int { return stream_.device_id(); }
 
   auto forward() -> void override {
+#ifdef SPFFT_ROCM
+    // workaround for bug with rocFFT for case 1x1xZ
+    if (spaceDomain_.dim_mid() == 1 && spaceDomain_.dim_inner() == 1) {
+      // make sure imaginary part is 0
+      gpu::check_status(gpu::memset_async(
+          static_cast<void*>(freqDomain_.data()), 0,
+          freqDomain_.size() * sizeof(typename decltype(freqDomain_)::ValueType), stream_.get()));
+      // copy real valued data into complex buffer -> from stride 1 to stride 2
+      gpu::check_status(gpu::memcpy_2d_async(static_cast<void*>(freqDomain_.data()), 2 * sizeof(T),
+                                             static_cast<const void*>(spaceDomain_.data()),
+                                             sizeof(T), sizeof(T), freqDomain_.dim_outer(),
+                                             gpu::flag::MemcpyDeviceToDevice, stream_.get()));
+      // no transform needed
+      return;
+    }
+#endif
+
     if(symm_) {
       // Make sure buffer is zero before transform, such that the symmtry operation can identify
       // elements, which have not been written to by the FFT
@@ -213,6 +230,19 @@ class TransformReal2DGPU : public TransformGPU {
   }
 
   auto backward() -> void override {
+#ifdef SPFFT_ROCM
+    // workaround for bug with rocFFT for case 1x1xZ
+    if (spaceDomain_.dim_mid() == 1 && spaceDomain_.dim_inner() == 1) {
+      // copy complex data into real valued buffer -> from stride 2 to stride 1
+      gpu::check_status(gpu::memcpy_2d_async(static_cast<void*>(spaceDomain_.data()), sizeof(T),
+                                             static_cast<const void*>(freqDomain_.data()),
+                                             2 * sizeof(T), sizeof(T), freqDomain_.dim_outer(),
+                                             gpu::flag::MemcpyDeviceToDevice, stream_.get()));
+      // no transform needed
+      return;
+    }
+#endif
+
     gpu::fft::check_result(gpu::fft::set_work_area(planBackward_, workBuffer_->data()));
     gpu::fft::check_result(
         gpu::fft::execute(planBackward_, freqDomain_.data(), spaceDomain_.data()));

diff --git a/src/gpu_util/gpu_pointer_translation.hpp b/src/gpu_util/gpu_pointer_translation.hpp
@@ -56,7 +56,13 @@ auto translate_gpu_pointer(const T* inputPointer) -> std::pair<const T*, const T
     const T* devicePtr = nullptr;
     return {inputPointer, devicePtr};
   } else {
+#ifdef SPFFT_ROCM
+    // Bug with HIP (ROCm 3.5) causes attr.devicePointer to be set to the start of allocated memory
+    // -> pointers with offset are not correctly returned
+    return {static_cast<const T*>(nullptr), inputPointer};
+#else
     return {static_cast<const T*>(attr.hostPointer), static_cast<const T*>(attr.devicePointer)};
+#endif
   }
 }
 

diff --git a/src/gpu_util/gpu_runtime_api.hpp b/src/gpu_util/gpu_runtime_api.hpp
@@ -99,6 +99,7 @@ constexpr auto MemoryTypeManaged = GPU_PREFIX(MemoryTypeManaged);
 
 constexpr auto MemcpyHostToDevice = GPU_PREFIX(MemcpyHostToDevice);
 constexpr auto MemcpyDeviceToHost = GPU_PREFIX(MemcpyDeviceToHost);
+constexpr auto MemcpyDeviceToDevice = GPU_PREFIX(MemcpyDeviceToDevice);
 
 constexpr auto EventDefault = GPU_PREFIX(EventDefault);
 constexpr auto EventBlockingSync = GPU_PREFIX(EventBlockingSync);
@@ -179,11 +180,21 @@ inline auto memcpy(ARGS&&... args) -> StatusType {
   return GPU_PREFIX(Memcpy)(std::forward<ARGS>(args)...);
 }
 
+template <typename... ARGS>
+inline auto memcpy_2d(ARGS&&... args) -> StatusType {
+  return GPU_PREFIX(Memcpy2D)(std::forward<ARGS>(args)...);
+}
+
 template <typename... ARGS>
 inline auto memcpy_async(ARGS&&... args) -> StatusType {
   return GPU_PREFIX(MemcpyAsync)(std::forward<ARGS>(args)...);
 }
 
+template <typename... ARGS>
+inline auto memcpy_2d_async(ARGS&&... args) -> StatusType {
+  return GPU_PREFIX(Memcpy2DAsync)(std::forward<ARGS>(args)...);
+}
+
 template <typename... ARGS>
 inline auto get_device(ARGS&&... args) -> StatusType {
   return GPU_PREFIX(GetDevice)(std::forward<ARGS>(args)...);